Bayesian Population Pharmacokinetic/Pharmacodynamic

Modeling to Study the Effect on the Cardiovascular

Syndrome of the QTc Interval Prolongation of
Non-antiarrhythmic Drugs
Giovanni Smania
July 15, 2013

Alla mia intera famiglia,

fonte inesauribile di motivazione e supporto.

Contents
1 Introduction
1.1 Physiolgy of the QT Interval and the Drug Effect on it .
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Basics of Non Linear Mixed Effect Methods . . . . . . .
1.3.1 The Basic Model . . . . . . . . . . . . . . . . . .
1.3.2 Pharmacokinetic Application of a NLME Model
1.4 Basics of Bayesian Estimation . . . . . . . . . . . . . . .
2 Materials and Methods
2.1 Preclinical Protocol . . . . . . . . . . . . . . . . . . .
2.1.1 Pharmacokinetic Data . . . . . . . . . . . . .
2.1.2 Pharmacodynamic Data . . . . . . . . . . . .
2.2 Clinical Protocol . . . . . . . . . . . . . . . . . . . .
2.2.1 Pharmacokinetic-Pharmacodynamic (PK-PD)
2.3 Pharmacokinetic Modelling . . . . . . . . . . . . . .
2.3.1 NONMEM Model Formulation . . . . . . . .
2.4 PK-PD Modeling of the QT interval . . . . . . . . .
2.4.1 Modelling the Drug Effect . . . . . . . . . . .
2.4.2 Modelling the QT-RR Relationship . . . . . .
2.4.3 Modelling the Circadian Rhythm . . . . . . .
2.4.4 WinBUGS Model Formulation . . . . . . . .
3 Results
3.1 Pharmacokinetic Analysis .
3.1.1 Preclinical Data . .
3.2 Pharmacodynamic Analysis
3.2.1 MCMC Convergence
3.2.2 Preclinical Data . .
3.2.3 Clinical Data . . . .

. . . .
. . . .
. . . .
Check
. . . .
. . . .

.
.
.
.
.
.

4 Conclusions

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

. . . .
. . . .
. . . .
. . . .
Data
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

1
1
2
4
4
6
9

.
.
.
.
.
.
.
.
.
.
.
.

13
13
13
13
14
14
14
15
15
16
17
18
19

.
.
.
.
.
.

27
27
27
33
33
36
39
43

iii

Abstract
Assessment of the propensity of non-antiarrhythmic drugs in prolonging QT/QTc interval
is critical for the progression of compounds into clinical development. Given the
similarities in QTc response between dogs and humans, dogs are often used in pre-clinical
cardiovascular safety studies. The current regulatory guidelines are based on simple
statistical analyses of QT data, thereby ignoring any potential exposure-effect relationship
and nonlinearity in the underlying physiological fluctuation in QT values. The objective
of this analysis is to adopt a model based approach to assess the QT-prolonging effects in
dogs and humans of GSK945237, a new compound under development. Pharmacokinetic
and pharmacodynamic data from experiments in conscious dogs and clinical trials
following administration of GSK945237 were used. First, pharmacokinetic modelling was
applied to derive drug concentrations at the time of each QT measurement. A Bayesian
PKPD model was then used to describe QT prolongation, which allows discrimination of
drug-specific effects from other physiological factors known to alter QT interval duration.
Results from this analysis show the drug under investigation is not prone to cause
hazardous increases in the QT interval for both humans and dogs. Further, the PKPD
model is capable to predict both preclinical and clincal data, suggesting it might be used
for future translational research in the field of QT prolongation.

Chapter 1

Introduction
1.1

Physiolgy of the QT Interval and the Drug Effect on it

The QT interval is the portion of the Electrocardiographic signal (ECG) in between the
beginning of the QRS complex and the end of the T-wave (Figure 1.1). It represents the
time between the onset of electrical depolarization of the ventricles and the end of their
repolarization, that is, it reflects the duration of individual action potentials in the cardiac
myocytes.
QT interval is affected by several sources of variability of diverse nature. As one would
excpect, the QT interval duration is strictly related to the RR interval (i.e. the cardiac
period). In fact it is well known that in order to improve the detections of threatening QT
prolongations, the measured QT has to be corrected for changes in RR interval, taking the
name of corrected QT interval (QTc). Other subject-related factors potentially affecting
the QT/QTc interval are:
1. Genetic (long QT syndrome)
2. Food intake
3. Circadian rhythm
4. Sex
5. Obesity
6. Physical activity
7. Blood glucose level
8. Blood pressure
9. Age
1

CHAPTER 1. INTRODUCTION

Figure 1.1: Schematic ECG trace showing the different waves and intervals that characterize it.
In cardiovascular safety assessment studies, considerable attention has been paid due
to the fact that both antyarrhythmic and non-antyarrhythmic drugs may prolong the QT
interval.
QT prolonging compounds do that by blocking the so-called human Ether-
a-go-go Related
Gene (hERG), a gene coding for the alpha subunit of a potassium ion channel (also called,
with a bit of ambiguity, hERG) which modulates the repolarizing current IKr in the cardiac
action potential.
A possibile explanation for such drug effect has been given by Milnes et al. [1]. Starting
from the hypothesis that the high-affinity drug-binding sites are within the vestibule of the
channel, they state that the absence of highly conserved prolines (normally present in K+
channels) at hERG ammino-acid positions may suggest that a kink is absent, thereby
more space is available within the vestibule. A potential effect of this increasing in space
inside the vestibule of the hERG would be the ability of a wide variety of drugs to gain
access to binding sites inside the pore.
The blockade of hERG channel implies a lower conductivity of K+ ions (i.e. a lower IKr ),
whom reflects on a bigger duration of the action potential of cardiac myocytes, ending with
a slower ventricular repolarization and a longer QT interval (Figure 1.2).

1.2

Motivation

Non clinical and clinical cardiovascular toxicities are one of the main safety reasons for
(i) drug discontinuation throughout all stages of drug discovery and development;

1.2. MOTIVATION

Figure 1.2: hERG blockade proccess.

(ii) serious adverse events and adverse drug reactions during clinical development and
(iii) post marketing drug withdrawal.
Among those, prolongation of the QT/QTc interval had been the reason for one-third of
all drug withdrawals for the period 1990-2006 [2].
Although drug-induced lengthening of the QT/QTc interval is not a safety concern in itself,
this effect has been associated with the occurrence of a rare, life-threatening ventricular
tachycardia named Torsades de Pointes (TdP) [3]. Therefore the ability to assess QT/QTc
prolongation effects as early as possible during the course of the clinical development of
a new compound is helpful not only in understanding the potential risk-benefit ratio of
the molecule but also in deciding whether to progress with the development of the drug,
thereby avoiding unnecessary costs.
According to the most current regulatory guideline - the International Conference of Harmonisation (ICH) E14 Document [4] - the critical threshold for cardiovascular safety concerns in the QTc-interval increasing is equal to 10 msec. Therefore, in the context of QTc
prolongation, the application of translational methods makes possible the prediction of
QTc lenghtening in humans starting from outcomes obtainend in preclinical studies; with
regard to this, the work that has been proposed provides a tool to investigate such translation and to identify the gap between preclinical and clinical findings. Further, given that
the protocol suggested by the ICH E14 can lead to biased results for several reasons [5], in
order to increase the reliability of potential conclusions the QT interval has been described
by means of nonlinear mixed-effects modeling techniques applied in a bayesian framework.
Non linear mixed-effect modeling helps to overcome issuses due to managing sparse data

CHAPTER 1. INTRODUCTION

coming from different individuals (which is the case for QT measurements in humans),
while a bayesian approach fits naturally into decision analysis since it allows to obtain the
full posterior probability plus the credible intervals of a parameter of interest instead of
a puntual estimate with its confidence interval. These two aspects will be treated widely
later (see 1.3 and 1.4).

1.3

Basics of Non Linear Mixed Effect Methods

Non Linear Mixed Effect (NLME) models, also called hirerchical models, are used in population analysis related fields such as environmental health, agriculture and pharmacokinetics/pharmacodynamics.
Population analysis is an alternative method to individual analysis that has its strength
in extrapolating information from sparse data coming from different individuals, assuming
that each individual feature comes from a distribution of that feature which is representative of the entire population of interest.
The mathematical formulation of the model will be presented (1.3.1), followed by an applied
pharmacokinetic example (1.3.2).

1.3.1

The Basic Model

NLME modelling is more complex than standard modelling techniques, and has to be
stratified into two hirerchical steps.
1. Stage 1: Individual-Level model
The individual step aims to define the measure model:
zi,j : the j th measure (j = 1, ..., N ) for the ith subject (i = 1, ..., K), collected at
time ti,j
ui : vector of additional conditions under which subject i is observed
pi : vector of model parameters specific to subject i (M 1)
f (ti,j , ui , pi ): non-linear relationship between zi,j and (ti,j , ui , pi )
i,j : random error due to uncertainty in the measure denoting within-individual
variabilty

If we consider an additive residual error, the individual model can be written as

zi,j = f (ti,j , ui , pi ) + i,j .

(1.1)

1.3. BASICS OF NON LINEAR MIXED EFFECT METHODS

Defining

i,1
f (ti,1 , ui , pi )
zi,1
i,2
f (ti,2 , ui , pi )
zi,2

i = .
zi = . fi (pi ) =

.
.
.
.

.
.
i,N
f (ti,N , ui , pi )
zi,N

it is possible to summurize equation (1.1) as

zi = fi (pi ) + i

(1.2)

where a classical assumption on i is

i |pi MVN (0, Ri (pi , ))
with costant accros individuals. For example it might need to incorporate a proportional error model, therefore will be constant (= 2 ) and Ri will be

f (ti,1 , ui , pi )
0

0
f (ti,2 , ui , pi )
0

Ri = 2 fi (pi ) = 2

..
..
.
.
.
.

.
.
.
.
0
0
f (ti,N , ui , pi )
2. Stage 2: Population-Level model
To account for inter-individual variation of parameters pi among individuals, a model
for pi has to be specified. The population step provides this model:
ai : vector of characteristics of subject i (named covariates)
i : vector of random effects for subject i conveying inter-individual variation
: vector of fixed effects expressing features common to the entire population
d: M-dimensional vector function

Then a general model for pi is given by

pi = d(ai , i , )
where a classical assumption on i is
i MVN (0, )

(1.3)

CHAPTER 1. INTRODUCTION

1.3.2

Pharmacokinetic Application of a NLME Model

Pharmacokinetics is the study of the time course of drug concentration in different body
spaces such as plasma, blood, urine, cerebrospinal fluid, and tissues, and aims to describe
the different phases encountered by the compound once in the body: absorption, distribution, metabolism and elimination. The representation of the body is approximated by
simple compartment models, each describing one body spaces, with the kinetics of the drug
described by differential equations.
Suppose data from N = 6 subjects is available with the drug given orally. Figure 1.3
shows the concetration-time profiles of the different individuals: it can be seen the similiraty in shapes across subjects, but peaks, rises, decays vary considerably. This effect
is attributable to inter-subject variation in underlying PK processes. Assuming a OneCompartment model with oral dose D (see Figure 1.4) given at time t = 0 leads to describe
concentration C(t) at time t 0 as
C(t) =

h Cl
i
Dka
V t
ka t
e

e
V (ka Cl
V )

where
ka : fractional rate of absorption [1/Hr] of the drug from the gut compartment
Cl: clearence rate [ml/Hr], i.e. the volume of plasma which per unit of time is totally
cleared of a substance (e.g. a drug) by the various elimination processes (metabolism
and excretion)
V : volume of distribution [ml], i.e. the theoretical volume that a drug would have
to occupy (if it was uniformly distributed), to provide the same concentration as it
currently is in blood plasma

summarize the PK processes underlying the observed concentration profile of a given subject.
The final goal of a PK analysis is to determine the population mean values of (ka , Cl, V )
and how they vary between subjects, elucidating whether part of this variation is associated with subject characteristics (i.e. covariates, such as weight, age, renal function), in
order to develop dosing strategies for subpopulations with certain characteristics (elderly,
pediatric, etc.).
With respect to the model formulation encountered in 1.3.1, a correlation with the quantities in play leads to
1. Individual Model
zi,j is the drug concentration for subject i at time ti,j , that is zi,j = C(ti,j )

1.3. BASICS OF NON LINEAR MIXED EFFECT METHODS

Drug Concentration [mg/l]

600
500
400
300
200
100
0
0

10

20

30

40

50

60

Time [Hr]
Figure 1.3: Example of drug concentration profiles after oral administration (each color
identifies a single subject).
ui is the dose given to subject i a time t = 0, that is ui = Di
pi are the PK parameters (ka , Cl, V ) specific to subject i, i.e. pi = (kai , Cli , Vi )
f (ti , ui , pi ) describes how the concentration of the ith subject evolves in time in
a one-compartment model with first order absorption, i.e.

 Cl

Di kai
V i ti
k
t
a
i
f (ti , ui , pi ) =
e i e i
i
Vi (kai Cl
Vi )
2. Population Model
T

ai = Gi BWi , where Gi and BWi are the gender and the body weight of
subject i respectively

CHAPTER 1. INTRODUCTION

Figure 1.4: One compartment model with first order absorption from the gut compartment.
Such model allows to take into account oral administration.

T
i = 1,i 2,i 3,i
2

2
2
13
11 12
2
2
2
22
23
i N (0, ), = 21
2
2
2
31 32 33
where diag() quantifies the varaibaility accross individuals of (ka , Cl, V ), respectively.

T
= 1 2 3 4
The d function is the one that relates (kai , Cli , Vi ) to (ai , i , ), for example:

kai = 1 e1
Cli = (2 + Gi 3 ) e2
Vi = (BWi 4 ) e3 .
In the pharmacokinetic context the use of exponentials on random effects i is required
becuase of dealing with physiological entities, which are positive by implication. In addition, in the formulas above, an influence of the gender on the clearance and an influence
of the body weight on the volume of distribution is assumed.

1.4. BASICS OF BAYESIAN ESTIMATION

1.4

Basics of Bayesian Estimation

In the branch of estimation theory, Bayesian approach refers to the situation in which the
information available comes from two different sources. Defining y as the observed data
and as the quantity to be estimated (a vector of parameters, a signal, etc.), these two
founts are:
Likelihood: the knowledge on the relationship between y and , denoted as p(y|);
Prior: the probabilistic a priori information on , denoted as p();

thus, both y and are treated as random variables.

Bayes theorem (equation (1.4)) enables to combine these two probability densities in order
to obtain the so-called posterior density function p(|y):
p(|y) =

p(y|)p()
p(y)

(1.4)

Translating the theorem in practical term: p() represents the prior knowledge we have
on the quantity to be estimated, if such quantity is somehow related to data y, the observation of y (i.e. p(y|)) changes our expectations on in p(|y). In other words, by
applying Bayes theorem we are extracting the estimate by way of a compromise between
our knowledge on the quantity to be estimated and what the data is telling us (see Figure
1.5).
It is easy now to deduce the main difference between Bayesian inference and other approaches: through Bayes theorem the entire distribution of the parameters of interest can
be estimated, providing a great amount of information. In fact, the use of posterior distribution is indeed wide, both because there are different ways to obtain the point-estimator
of the parameter (such as the posterior mean, the posterior median or the posterior mode)
and because it provides the credible intervals, which are a reliable measure (compared to
confidence intervals used in frequentist approach) of the uncertainty of an estimate.
In order to be able to evaluate the postierior density function, Bayes theorem of equation
(1.4) can be rewritten, using the Law of Total Probability, as
p(y|)p()
.
p(y|)p()d

p(|y) = R

(1.5)

The denominator of equation (1.5) is a constant and does not depend on ; hence, in
the intereset of simplifying the formula as much as possible, we may consider
p(|y) p(y|)p()

(1.6)

Unfortunately, only in a few particular cases p(|y) is available in closed form, that is
when the prior and the posterior come from the same family of distributions, and in this

10

CHAPTER 1. INTRODUCTION

case the prior is said to be conjugate to the likelihood. Example 1.4.1 shows one of these
cases.
Example 1.4.1. Let us consider a drug to be given for relief of chronic pain whose experience with similar compounds has suggested that annual response rates between 0.2 and
0.6 could be feasible, then we may interpret this as a distribution with mean = 0.4 and
standard deviation = 0.1. Suppose we now treat n = 20 volunteers with the compund and
observe y = 15 positive responses.
Likelihood. Assuming patients are independent, with common unknown response rate
, leads to a binomial likelihood (Figure 1.5a):

y Binomial(, n)
 
n y
p(y|n, ) =
(1 )ny y (1 )ny
y
Prior. A probability distribution which fits with the requests on the response rate is
a Beta distribution (with a=9.2 and b=13.8, see Figure 1.5b):

Beta(a, b)
p() =

(a + b) a1
(1 )b1 a1 (1 )b1
(a)(b)

Posterior. Combining the Binomial likelihood and the Beta prior gives the following
posterior distribution
p(|y, n) p(y|, n)p()

p(|y, n) y (1 )ny a1 (1 )b1 = y+a1 (1 )ny+b1

which is still a Beta dsitribution (with different parameters, see Figure 1.5c):
p(|y, n) Beta(y + a, n y + b)

In most cases p(|y) is not analitically tractable and we want to obtain theRmarginal posterior p(i |y) for each
R i and calculate properties of p(i |y) such as the mean ( i p(i |y)di )
and the tail areas ( T p(i |y)di ). Given that evaluating these integrals analitically is impossible, numerical integration becomes vital. WinBUGS [6] is a free software that allows
to obtain every p(i |y) and compute integrals on them via Markow Chain Monte Carlo
(MCMC) techniques.

1.4. BASICS OF BAYESIAN ESTIMATION

11

0.20

pdf(y|)

0.15

0.10

0.05

0.00
0.0

0.2

0.4

0.6

0.8

1.0

(a)

pdf()

0
0.0

0.2

0.4

0.6

0.8

1.0

(b)

0.05

pdf(|y)

0.04

0.03

0.02

0.01

0.00
0.0

0.2

0.4

0.6

0.8

1.0

(c)

Figure 1.5: Probability density function (up to a multiplicative constant) of: Likelihood (a),
Prior (b) and Posterior (c) defined in example 1.4.1. It may be noticed how the posterior
obtained would be in the middle (a compromise) between the prior and the likelihood.

Chapter 2

Materials and Methods

2.1

Preclinical Protocol

2.1.1

Pharmacokinetic Data

The compound (GSK945237) was administered orally twice daily (Bis In Die, i.e. BID)
with approximately 6 hours between doses to nine male and nine female beagle dogs (total
number of individuals=18) at total daily doses of 0, 30, 100 and 300 mg/kg/day (0, 15,
50 and 150 mg/kg/day BID, respectively). Plasma samples were obtained from blood
collected into tubes containing EDTA from drug-treated and placebo-control animals at
the following nominal times: predose, 0.25, 0.5, 1, 3, 6, 6.25, 6.5, 7, 9 and 24 hours after
the first dose.
Plasma samples were analyzed for GSK945237 using a validated analytical method based
on protein precipitation, followed by HPLC-MS/MS analysis. Using a 50 L aliquot of
dog plasma, the lower limit of quantification (LLQ) for GSK945237 was 0.1 ng/L and the
higher limit of quantification (HLQ) was 50 ng/L.

2.1.2

Pharmacodynamic Data

Four male dogs were given placebo and total daily doses of 30, 100 and 300 mg/kg (15, 50
and 150 mg/kg administered twice daily approximately 6 hours apart) of test article orally
by gavage on separate days, with at least 7 days between each dose, according to a 4x4
latin square crossover paradigm. Dogs were dosed at approximately 9:00 AM on each day
of dosing. ECG waveforms were recorded continuously from 2 hours prior to dosing to 24
hours after dosing and all derived parameters were recorded as 1-minute means.
According to the aim of the analysis and given the enormous amount of data collected
overall, raw QT data was diluted every 2 minutes during the absorption phase (roughly
t<2.5hrs and 6hrs<t<8.3hrs, where t is the time after dosing), every 5 minutes during
the distribution phase (2.3hrs<t<6hrs and 8.3hrs<t<15hrs) and every 15 minutes during
13

14

CHAPTER 2. MATERIALS AND METHODS

the elimination phase (t>15hrs). Also, part of the ECG data collected around the dosing
times had to be discarded due to the bad quality of the data caused by an abnormal dogs
excitement.

2.2
2.2.1

Clinical Protocol
Pharmacokinetic-Pharmacodynamic (PK-PD) Data

The study was a first-time-in-human to investigate the safety, tolerability, and pharmacokinetics of escalating single oral doses of GSK945237. The adopted design was a single-blind,
randomized, placebo-controlled, dose-rising study conducted in sequential cohorts. Fortyfive (45) subjects were enrolled into 6 cohorts. According to a randomization schedule prepared prior to study start, 6 subjects received active drug and 3 subjects received placebo
at each dose level (50mg, 150mg, 400mg, 1000mg, 2000mg and 2600mg). Each subject
participated in one study period and received either GSK945237 or matching placebo.
Following a single oral dose of GSK945237, blood samples for pharmacokinetic analysis
and ECG waveform samples for pharmacodynamic analysis of GSK945237 were collected
over a 60-hour period from groups receiving doses of 50mg, 250mg and 500mg and over
a 120-hour period from groups receiving 1000mg and 1750mg. Bioanalysis of GSK945237
plasma concentrations was conducted using a validated analytical method based on solid
phase extraction followed by HPLC-MS/MS. The lower limit of quantification (LLQ) for
GSK945237 was 10 ng/mL, using a 100 L aliquot of human plasma with a higher limit of
quantification (HLQ) of 5000 ng/mL.

2.3

Pharmacokinetic Modelling

The goal of the pharmacokinetic (PK) modelling in the whole project is to obtain time
matched concentration and QT interval values, since this is required for the assessment
of the pharmacokinetic-pharmacodynamic (PK-PD) relationship. In order to do this, PK
modelling has been performed using non-linear mixed effect techniques (see 1.3.2 for an
example) implemented in NONMEM 7.1.2 [7]. Once identified the best PK model for
GSK945237, the parameter estimates obtained are used to generate simulated concentration profiles with the same sampling time used to collect QT measurements.
Since in human data QT measurements and concentration measurements were collected
simultaneously, the PK step was executed only for conscious dogs data, where the QT
sampling time is usually low (30 sec/1 min) and therefore an equal PK sampling time is
not achievable for obvious reasons.
Firstly, provided that the drug administration was oral, a model with first order absorption
had to be taken into account. Secondly, the BID way of administration made it difficult to
describe the data after the second daily dose since the dogs had free access to food between

2.4. PK-PD MODELING OF THE QT INTERVAL

15

the two daily doses, determining the appearaence of a food effect in the data. Finally, the
NONMEM analysis led to the conclusion that a one compartment model with first order
absorption (Fig.1.4) best decribes the kinetic of the analyzed drug.

2.3.1

NONMEM Model Formulation

The mandatory parameters for a one compartment model with first order absorption are:
Rate of Absorption (Ka ), Clearance (Cl) and Volume of Distribution (V ). Bioavailability
(F ) was also included in the model in order to describe the food effect.
After a covariate analysis, it emerged that the Clearance and the Volume of Distribution
are affected by the gender and the body weight of the individuals, respectively.
The detailed model formulation is (t represents the time variable):

F ECl =

1
2

if GENDER=M
= Cl = F ECl e1
if GENDER=F
V = W T 3 e2


F EKa =

4
5

7
F =

8
1

if First Daily Dose (FDD)

= Ka = F EKa e3
if Second Daily Dose (SDD)
if DOSE=30 mg/kg and DOSE=SDD
if DOSE=300 mg/kg and DOSE=FDD
if DOSE=300 mg/kg and DOSE=SDD
otherwise

(2.1)

(2.2)

The residual error model that has produced the lowest objective function value is a
mixed (proportional and addittive) one:
Y = f (, , t) (1 + 1 ) + 2

(2.3)

where Y is the observed concentration, f (, , t) the predicted concentration, 1

N (0, 12 ) and 2 N (0, 22 ).

2.4

PK-PD Modeling of the QT interval

The aim of this section is to define a mathematical model able to decribe the QT interval.
Since there are different sources contributing on QT variation, each of them is explored.

16

CHAPTER 2. MATERIALS AND METHODS

2.4.1

Modelling the Drug Effect

Pharmacodynamics can be defined as the relationship between the drug concentration

in the plasma and the pharmacological effect. In order to be able to characterize this
relationship and quantify the drug response, various mathematical models are available
[8]. Among those, the one that has been implemented is a linear concentration-effect
model (Figure 2.1)
E = E0 + S C
(2.4)
where E indicates the drug effect, E0 the baseline effect, C the plasmatic concentration of
the drug and S the slope parameter.

40

Effect [a.u.]

30

20

10

20

40

60

80

Concentration [a.u.]
Figure 2.1: Observed effect (dots) versus concentration described by a linear model (solid
line).
Despite an ordinary Emax model (Figure 2.2)
E = E0 +

Emax C
EC50 C

(2.5)

(where EC50 denotes the plasma concentration corresponding to the half-maximal difference between baseline E0 and the maximal effect Emax ) might be more physiologically
plausible as it takes into account the maximum effect, the range of concentration values
involved in QT studies does not determine any saturation effects, allowing the use of a

2.4. PK-PD MODELING OF THE QT INTERVAL

17

linear model.
In the context of QT variation, equation (2.4) can be rewritten as
QTDE = Slope C

(2.6)

where QTDE represents the variation of the QT interval due to the drug effect and
the unit of measurement of the Slope will be [ms/concentration]. The baseline effect E0
will be consider in section 2.4.2.

Effect [a.u.]

80

60

20

40

40

20

60

80

Concentration [a.u.]

Figure 2.2: Observed effect (dots) versus concentration described by an Emax model (solid
line).

2.4.2

Modelling the QT-RR Relationship

It has been already mentioned that the QT interval depends on the RR one, but how are
they related? Since 1920 when two pioneering article appeared [9],[10], many efforts have
been made to find a formula that provides the ideal correction such that the corrected QT
interval is indipendent of the RR interval.
The most used formulas so far are the so-called Bazett correction [9]
QT cB = QT RR0.5

(2.7)

18

CHAPTER 2. MATERIALS AND METHODS

and Friedericia correction [10]

QT cF = QT RR0.33

(2.8)

where QT is in [ms] and RR is in [s] (there is no a scientific reason why RR and QT

have two different measurement units in these formulas, on the contrary, it came out from
empirical evidence). Figure 2.3 compares the two formulas using as an example a pooled
dataset comprising QT-RR measurements from 45 individuals. Between the two formulas,
that of Friedericia (Figure 2.3b) performs on avarage quite well (the loess smoother is almost horizontal), while Bazett formula (Figure 2.3a) determines an over-correction.
Despite the good behaviour of Friedericias formula, equations (2.7) and (2.8) have been
derived from population data, and the large difference between these formulae suggests
that the QT/RR relationship has not been found reproducible from study to study. It is
therefore unreasonable to expect that a general formula could satisfy the QT/RR relationship for the data of a given study, namely, individuals show different QT/RR relationships.
Keeping the same template of equations (2.7) and (2.8), the QTc baseline interval (QT c0 )
can be described as
QT c0 = QT0 RR
(2.9)
where QT0 is the measured QT baseline interval. Now, given that the aim is to describe
the QT interval, equation (2.9) has to be rewritten as
QT0 = QT c0 RR

(2.10)

A population approach makes possible to estimate every individual correction factor i

obtaining a population parameter denoting the QT correction factor for the population
recruited in the specific study. Furthermore, QT0 in equation (2.10) plays the role of the
baseline effect E0 introduced in equation (2.4) which was lacking in equation (2.6).

2.4.3

Modelling the Circadian Rhythm

As mentioned in section 1.1, the QT interval depends on several factors. Obviuosly, trying
to build a model able to describe each of them would be too complex as well as without
any pharmacological interest. Amongst the elements influencing the QT, apart from the
aforementioned RR interval and drug effect, the circadian rythm is one of the most sizable
(see Figure 2.4).
In general, the statistical significance of the circadian rythmicity can be documented
by the cosinor analysis [11], [12]. This method characterizes a rhythm by the parameters
of a fitted cosine function


QTcirc

2
= A cos
(tc )
24


(2.11)

2.4. PK-PD MODELING OF THE QT INTERVAL

19

where QTcirc denotes the variation in milliseconds of the QT interval due to the
circadian rythm, A [ms] and [hr] are its amplitude and phase respectively and tc is the
clock time.
The ability to estimate the crircadian component is limited by the number of QT interval
measurements available and their spread over the day. In fact, since the precision of QT
data is low, a large number of observations is needed to extract a signal coming from the
circadian ryhtm.
Concluding, the final model may be obtained by gathering each of the single contributes
described so far, i.e. the drug effect (equation (2.6)), the QT/RR relationship (equation
(2.10)) and the circadian rythm (equation (2.11)), that is, QT = QT0 + QTcirc + QTDE ,
which leads to


2
QT = QT c0 RR + A cos
(tc ) + Slope C
(2.12)
24

2.4.4

WinBUGS Model Formulation

The PK-PD model of equation (2.12) has been implemented in WinBUGS via a hierarchical
model which allows to take into account different sources of variability, as well as prior
distributions (Figure 2.5).
The detalied formulation is the following (hereafter: i = 1, ..., N individuals, k = 1, ..., P
occasions and j = 1, ..., M time samples)
QTi,k,j = fk,j (pi ) + j

(2.13)

where j is the measurement error, at time j, assumed normally distributed (j

N (0, 2 )) and


2
i
fk,j (pi ) = QT c0 i,k RRi,j
+ Ai cos
(tcj i ) + Slopei Ci,j
(2.14)
24
is the direct translation of equation (2.12).
Carrying on with the fromulation, pi is a vector of individual parameters

T
pi = p1,i p2,i p3,i p4,i p5,i
which will be used to derive the actual parameters appearing in the model and follows
a multivariate normal distribution (MVN)
pi MVN (, )
where


T
= 1 2 3 4 5

20

CHAPTER 2. MATERIALS AND METHODS

represents the vector of the fixed-effects and

2
2
11 12
2
2
21 22
2
2
=
31 32
2
2

41 42
2
2
51
52

2
13
2
23
2
33
2
43
2
53

2
14
2
24
2
34
2
44
2
54

2
15
2
25

2
35

2
45
2
55

depicts the random-effect component, i.e. the Between Subject Variabilty (BSV) of the
parameters.
The parameters of equation (2.14) are expressed as follows:
QT c0i,k = epi,1 +k

(2.15)

where the variable k is normally distributed around 0 with variance 2 (i.e. k

N (0, 2 )) and has been introduced in order to take into account the variability of the
baseline QTc across study periods (the baseline QTc varies from day to day). Such a
source of variabilty is called Inter-Occasion Variability (IOV). Furthermore, as the QT c0 is
a physiological parameter and therefore can not be negative, we make sure it will assume
only positive values by taking the exponent of pi,1 + k .
i = epi,2
It has already been described how the RR affects the QT interval (see section 2.4.2).
In particular, from equation (2.10), we know that the parameter has to be positive. This
is the reason why the exponent of pi,2 was considered.
Ai = epi,3
i = epi,4
A and are the model parameters used to describe the circadian component of the QT
interval. Theoretically, the sign of these two parameters is arbitrary, therefore no positive
constrains should be adopted. However, from a practical point of view, i.e. in order to
narrow the optim parameters searching field, exponential terms were considered, forcing A
and to be positive. With the aim of being able to assert that this choice does not affect
the parameters estimate, a model run with a negative constrain on A was performed (i.e.
Ai = epi,3 ). Results shown in Figure 2.6 enable to draw the conclusion that the model
recognizes the fact that the amplitude cannot be positive, and in consequence shifts the
phase to a value 12Hr greater than the previous one, thereby obtaining the same overall
estimate of the cyrcadian rhythm component.
Slopei = 105 pi,5

(2.16)

2.4. PK-PD MODELING OF THE QT INTERVAL

21

The formulation of Slope showed above contains a multiplication by 105 , whose goal
comes from numerical issues. Since Slope parameter is known to be low, equation (2.16)
ensures that pi,5 will be a reasonable high number, supporting the algorithm in finding a
solution.
Priors Choice The selection of a prior for a given parameter is, in principle, subjective:
it might be elicited from experts, it might be based on historical data (e.g. a previous
study) or it might be convenient to assume a vague, non-informative prior.
As to priors on 2 and 2 , a Jeffreys prior on the inverse of these two quantities is a
standard choice if no a-priori information is present on them (non-informative prior):
1
Gamma(, )
2
1
Gamma(, )
2
where is an arbitrary small number (usually = 0.001).
Concerning the parameters and , the choice was also for non-informative priors in order
to do not bias the estimation too much. In particular, the priors were specified as:
MVN (, )
where

0
1
0
0

=
0 , = 10 0
0
0
0
0

0
1
0
0
0

0
0
1
0
0

0
0
0
1
0

0
0

0
1

and
1 Wp(R, 5)
where Wp(R, ) indicates a Wishart distribution with scale matrix R and degrees of
freedom and it is a common choice for prior of the inverse covariance-matrix (i.e. 1 )
of a multivariate-normal random-vector (i.e. ) since it is a conjugate prior for it [13]. To
represent vague prior knowledge, was set small as possible (i.e. 5, the rank of ). Finally,
the scale matrix was specified as

1 0 0 0 0
0 1 0 0 0

R=
0 0 1 0 0 ,
0 0 0 1 0
0 0 0 0 1

22

CHAPTER 2. MATERIALS AND METHODS

an arbitrary, as well as common, option given that the choice of R has little effect on
the posterior estimate of [14].

2.4. PK-PD MODELING OF THE QT INTERVAL

23

450

QTcB [ms]

400

350

0.6

0.8

1.0

1.2

1.4

RR [s]
(a)

460

440

QTcF [ms]

400

380

360

340

420

0.6

0.8

1.0

1.2

1.4

RR [s]
(b)

Figure 2.3: Comparision between the RR correction formulas of equations (2.7) and (2.8):
(a) Bazett correction and (b) Friedericia correction. The dots represent the QTc measurements. The loess smoother (solid line) shows a better performance for Fridericia correction.

24

CHAPTER 2. MATERIALS AND METHODS

10

12

14

16

18

440
420
400

QT [ms]

380

360
340

440

420

400

380

360

340

10

12

14

16

18

Clock Time [Hr]

Figure 2.4: QT intervals (circles) measured in placebo groups vs clock time. The solid line
is a loess smoother showing a circadian variation in the QT.

2.4. PK-PD MODELING OF THE QT INTERVAL

25

Figure 2.5: Graph of the hierarchical model used for the PK-PD analysis. The different
levels are, from lower to higher: Prios level, Population level, Individual level, Occasion
level and Measurement level.

CHAPTER 2. MATERIALS AND METHODS

Circadian Variation of QT interval [ms]

26

10
A=9.066ms =15.78hr
A=9.059ms =28.14hr

8
6
4
2
0
2
4
6
8
10

10

15

20

25

Clock Time [hr]

Figure 2.6: Comparison between the circadian rhythm estimated with a positive (solid
line) and negative (dashed line) constrain on the amplitude.

Chapter 3

Results
3.1
3.1.1

Pharmacokinetic Analysis
Preclinical Data

In this section, results from the preclinical PK model of GSK945237 (described in section
2.3.1) will be presented and discussed. The outcomes will be interpreted through both a
numerical and graphical point of view. All the estimates that are going to be documented
were obtained in NONMEM adopting the pre-implemented Maximum Likelihood estimation algorithm called First Order Conditional Estimation (FOCE) with interaction.
A fair diagnostic plot to see how the model works at first glance is the so called Goodness
of Fit (GOF) plot for population prediction (Figure 3.1). The population GOF plot aims
to show the overall behaviour of the model, and the key point to interpret it is that the
closer the black dots are to the unity line the better the prediction is, since this is proof
of a good correlation between observed and predicted concentrations. The plot reveals a
difficulty in getting the higher concentration values as there is a slight shift of the dots under the unity line for measured concentrations in between 23-30 ng/L. Nevertheless, given
the small number of indivduals involved in the study (18), the model may be accepted
and subjected to further examination. Figure 3.2 illustrates the GOF plot for individual
predicted concentrations, which are obtained by way of a Maximum a Posteriori (MAP)
Bayesian estimatation of the indivdual PK parameters using the population estimates as
priors (this step is automatically done by NONMEM). In this case the correlation between
predictions and observations is higher than in the population GOF, which is what one
would expect since there is an individual prediction for every subject.
The residuals analysis is an alternative way to evaluate the goodness of a model. A residual
is defined as the difference between the observed measure and the model prediction, therefore the lower the residuals are the better the prediction is; on the other hand, weighted
residuals are residuals scaled by a certain weight parameter (for example the variance of
the measurement error, if we have information on it) that quantifies the confidence we have
27

28

CHAPTER 3. RESULTS

on a certain sample. Under certain statistical assumptions, in order to assess the goodness
of the prediction, two particular features of the weighted residuals have to be checked: correlation (the more they are uncorrelated the better the fit is) and amplitude (they should
be distributed around -1 and 1). Outcomes from such analysis are showed in Figure 3.3,
and it can be seen the performance of the model is overall good, except for a sistematic
bias in the last sample. Given the goal of the PK step though, since we are interested in
higher concentrations and the model lacks on the last sample which is associated with a
very low concentration value, this flaw may be neglected.

Population Prediction [ng/L]

25

20

10

15

10

20

30

Observed Concentration [ng/L]

Figure 3.1: Goodness of Fit plot for the population preditced vs. observed drug concentration (the solid line depicts the unity line).
A more accurate diagnostic tool is represented by the Visual Predictive Check (VPC),
which is mainly used to support the appropriateness of a model. The VPC is constructed
from stochastic simulations of the model, therefore all the model components contribute
and it can help in diagnosing both structural and stochastic contributions. In Figure 3.4 is
portrayed the outcome from 1000 models simulations where the 25th and 75th percentiles
have been chosen due to the small size of the dataset [15]. The plot displays the capacity
of the model to express the variabilty that exists among individuals, where the accuracy in
doing that is confirmed by the fact that the measurements lying further the two percentiles
are around the 25% of the whole dataset (Table 3.1).
Now that the goodness of the model has been established, all the parameter estimates
may be explored.

3.1. PHARMACOKINETIC ANALYSIS

29

Individual Prediction [ng/L]

30

20

10

10

20

30

Observed Concentration [ng/L]

Figure 3.2: Goodness of Fit plot for the individual preditced vs. observed drug concentration (the solid line depicts the unity line).
Percentage of data above the
75th percentile
28.8%

Percentage of data beneath the

25th percentile
24.7%

Table 3.1

Table 3.2 contains the estimates (with their coefficient of variations) of the fixed effects
of the model. Referring to Figure 3.5 and the values obtained for 4 and 5 , it should be
appear clear why the Ka has been formulated as in equation (2.1): mostly due to the food
effect, the rate of absorption of the first daily dose is rather higher than the one in the
second dose. Also, as it may seen from the large variability in the second peak between
Figures 3.5a, 3.5b and 3.5c, the fact that dogs were fed between the two daily doses has
an influence in the maximum concentration of the second daily dose as well, depending on
the amount of drug administered. In the interest of making the model sensitive to such
effect, three different bioavailabilities for the three different dose regimens (30, 100 and
300 mg/kg) were taken into account (equation (2.2)), where the full bioavailabilty (i.e.
F = 1) was assumed to be the one on the first daily dose and on the 100 mg/kg dosing
regimen.

30

CHAPTER 3. RESULTS

Population Weighted Residuals

10

15

20

Time [h]

(a)

Individual Weighted Residuals

1.0

0.5

0.0

0.5

1.0

10

15

20

Time [h]
(b)

Figure 3.3: Population weighted residuals (a) and individual weighted residuals (b).

3.1. PHARMACOKINETIC ANALYSIS

31

30

Concentration [ng/L]

20

10

10

15

20

25

Time [Hr]

Figure 3.4: Visual Predictive Check of the PK model (1000 simulations). Dots: observed
concentrations, solid line: median of the 1000 simulated profiles, dashed lines: 25th and
75th percentile of the 1000 simulated profiles.
1 
Value
CV (%)

6.34
10.1

L
hr

 2
4.51
8.8

3
[L]
3.45
7.1

4 
8.59
20.8

1
hr

 5

0.789
23.4

7
8
unitless
0.609 0.44 0.78
9.7
11 12.5

Table 3.2: Estimates of the population mean parameters along with their precisions (expressed via percentage coefficient of variation) of the PK model.

Finally, the estimates of the random effects are proposed. As discussed in section 1.3.1,
in the first place the random effects are associated with two sources of variability: Between
Subject (Table 3.3) and Intra (Within) Subject (Table 3.5).
The estimates of the BSV in Table 3.3 are depicted by means of coefficients of variation, that
is BSVn = nn , where n is the true enitity that expresses the BSV of the nth (n = 1, 2, 3)
parameter, but it is divided by the population mean in the interest of having a relative
measure of the variability of that particular parameter. A further indicator of the goodness
of the BSV estimates is the result of an hypothesis test on the values of n , assuming as
null hypothesis that n has mean zero (see equation (1.3)). Table 3.4 summarizes the outcomes of this test, showing p-values that allow to accept the null hypothesis, i.e. all the n
have mean zero, giving an additional evidence of the appropriateness of the proposed model.
Regarding the Intra Subject Variability (ISV) estimates of Table 3.5, the proportional

32

CHAPTER 3. RESULTS

Concentration [ng/L]

10

15

20

25

Time [Hr]

(a)

Concentration [ng/L]

10

10

15

20

25

Time [Hr]

(b)

30

Concentration [ng/L]

25

20

15

10

10

15

20

25

Time [Hr]

(c)

Figure 3.5: Obsereved concentrations (dots), population prediction (dashed line) and individual prediction (solid line) obtained from the PK analysis. Indivduals showed come from
the three dosing groups: (a) 30 mg/kg/day, (b) 100 mg/kg/day and (c) 300 mg/kg/day.

3.2. PHARMACODYNAMIC ANALYSIS

BSV (%)
CV (%)

33

Ka
FDD SDD
8.0
86.8
28.2 28.2

Cl
M
4.2
69.7

V
F
5.9
69.7

4.7
49.7

Table 3.3: Estimates of the Between Subject Variability (expressed via percentage coefficient of variation) along with their precisions (expressed via percentage coefficient of
variation) of the PK parameters.

Sample Average
p-value

1
1.24 103
0.98

2
9.13 103
0.76

3
1.37 103
0.93

Table 3.4: Sample averages accross individuals of the random effects. The P-values come
from a hypothesis test where the null hypothesis is that the mean of n is zero.
term of equation (2.3) was evaluated as 1 %, whilst the addittive term as 2 .

ISV
CV (%)

P roportional(1 )
23.4%
20.3

Addittive(2 )
0.144 ms
33.2

Table 3.5: Estimates of the Intra-Subject Variability (ISV) along with their precisions
(expressed via percentage coefficient of variation) described in equation (2.3).

3.2

Pharmacodynamic Analysis

As it was already mentioned in section 1.4, the amount of information that Bayesian
inference can provide is considerable and will be discussed in sections 3.2.2 and 3.2.3.
Further, performing such analysis in WinBUGS entails the necessity of checking whether
the different runs converged correctly (section 3.2.1).

3.2.1

MCMC Convergence Check

WinBUGS analysis is based on MCMC methods, a class of algorithms for sampling from
probability distributions based on constructing a Markov chain that has the desired distribution as its equilibrium distribution (see [6] for details). An example of two Markov
chains is shown in Figure 3.6, and a first step to assess the convergence is to visually check
that the two chains are uncorrelated and overlap, which is the case for Figure 3.6.

34

CHAPTER 3. RESULTS

405

QT0 [ms]

400

395

390

385

5000

10000

15000

20000

Iteration

Figure 3.6: Standard output of a WinBUGS run: two Markov chains of QT c0 parameter
in humans.
A problem with MCMC methods is that convergence cannot always be diagnosed as
clearly as in optimization methods. Indeed, there are several diagnostic tools that have
been proposed since GIBBS sampling became popular in Bayesian inference ([16], [17], [18],
[19]) and it may be very diffcult to satisfy all these criterions. In order to do not make the
convergence check step too complex and lenghty, only the Gelman-Rubin diagnostic [17]
will be performed, along with the assessment of low Monte Carlo (MC) errors, essential for
precise estimation of the posterior distribution [20].
Gelman and Rubin method constits in different steps. Firstly, to evaluate the within chain
variance W (hereafter j = 1, ..., m chains, i = 1, ..., n iterations and ij is the ith sample of
the j th chain of an arbitrary unknown parameter ) as
m

W =

1 X 2
sj
m
j=1

where s2j is the sample variance of in chain j:

n

s2j

1 X
(ij j )2
=
n1
i=1

and j is the sample mean of in chain j:

n

1X
j =
ij .
n
i=1

3.2. PHARMACODYNAMIC ANALYSIS

35

Secondly, to evaluate the between chain variance B as

m

B=

2
n X 
j
m1
j=1

where is the sample mean between chains:

m

1 X
=
j .
m
j=1

We can then estimate the variance of the stationary distribution as a weighted average
of W and B, i.e.

2 = (1

1
1
)W + B,
n
n

that is used to define the final parameter of interest, named Potential Scale Reduction
Factor (P RF )
s
P RF =

2
W

(3.1)

which should be as close to 1 as possible to state that the chains converged.

P RF values obtained from Markov chains of parameters of interest in both preclinical
and clinical data are presented in Table 3.6. The same table contains also index of
equation (3.2), which helps to assess the convergence achievement in terms of MC error;
particularly, should be less than 5%.
=

M Cerror
100
SDestimate

(3.2)

Table 3.6 suggests convergence of parameters A and in clinical data has not been
totally achieved, as indicated by a P RF > 1.5 and a > 5%. Nevertheless, increasing the
number of iterations does not solve the problem and the reason of a bad convergence most
likely lays in the low amount of QT samples (8) collected from each patients trhougout the
day. Moreover, in both species chains related to the drug effect parameter (i.e. Slope) also
present convergence issues, and this might be due to the fact that the parameter is moslty
zero (i.e. the drug effect is moslty absent) therefore chains struggle to overlap and follow
a straight pattern since sample values range in a very limited interval (see Figure 3.7).

36

CHAPTER 3. RESULTS

DOG
HUMAN
DOG
HUMAN

QT c0
1.00
1.00

PRF

1.01 1.05 1.14

1.48 2.95 1.89

Slope
2.21
1.17

0.92
1.49

(%)
0.98 1.29 2.65
4.76 5.57 5.17

6.11
5.69

Table 3.6: Potential Scale Reduction Factors and indices obtained after running the
WinBUGS model for both dog and human data.

Slope [ms/nM]

0.00010

0.00005

0.00000

0.00005

0.00010

2000

4000

6000

8000

Iteration

Figure 3.7: Markov chains of Slope parameter in preclinical data.

3.2.2

Preclinical Data

A thorough Bayesian analysis allows to estimate the full posterior distributions of the
investigated parameters, enabling to extrapolate quantitative information by means of
computing precentiles and means of such distributions. In regard to the PKPD model of
equation (2.12), Figure 3.8 illustrates the probabilty density function (pdf ) of parameters
QT c0 , , A, and Slope in conscious dog data. The shapes of these distributions supply
a qualitative idea on the range of values assumed by the parameters. Also, the shape acts
as an additional option of convergence inspection; the pdf, in fact, is an alternative way to
represent the run Markov chains for that parameter, thereby it should not contain peaks
and troughs since they would be evidence of lack of overlapping between chains as well as
correlation within chain.

3.2. PHARMACODYNAMIC ANALYSIS

37

0.12
0.10

pdf()

pdf(QTc0)

0.08
0.06

0.04
2

0.02
0.00

0
240

250

260

270

0.2

QTc0[ms]

0.3

(a)

0.4

0.5

(b)

0.20
0.10

0.08

pdf()

pdf(A)

0.15

0.10

0.06

0.04
0.05
0.02

0.00

0.00
0

10

15

20

25

30

20

40

60

[hr]

A [ms]

(c)

(d)

pdf(Slope)

15000

10000

5000

0
0.0001

0.0001

Slope [ms/Nm]

(e)

Figure 3.8: Posterior distributions of the PKPD model parameters in dog data: QT c0 (a),
(b), A (c), (d) and Slope (e).

38

CHAPTER 3. RESULTS

Posteriors obtained elucidates a successfull conclusion of the estimation step. Nonetheless, Figure 3.8e confirms what Table 3.6 had previoulsy told, i.e. a perfect convergence of
Slope Markov chains is difficult to achieve as elicited from the irregular shape of the pdf.

10% Credible Interval

Mean
90% Credible Interval
CV (%)

QT c0
[ms]
251.0
255.3
259.7
1.4

unitless
0.251
0.304
0.365
14.9

A
[ms]
6.34
8.77
11.93
28.1

[hr]
12.24
16.48
22.68
28.2

Slope
ms
[ nM
]
-1.2105
1.6105
4.5105
136.6

Table 3.7: Mean, 80% credible intervals and precision (expressed as percentage coefficient
of variation) of the population parameters of the PKPD model of equation (2.12) for dog
data.
A statistical analysis on the posteriors leads to results in Table 3.7, where the parameter
estimates are expressed in terms of posteriors mean (i.e. the minimum mean square error
(MMSE) estimator). Along with the MMSE estimator there are the 80% credible intervals
and the precision of the estimates expressed as percent coefficient of variation.
Once again, the low precision on Slope estimate (136.6%) is an indicator of the problems
arose in finding a complete convergence of the Markov chains for such parameter. An other
interesting consideration that can be done by looking at Table 3.7 concerns the negative
10% credible interval of Slope, which stresses the fact that the drug effect lays around zero
and therefore supports the conclusion that no drug induced QT prolongation takes place
in dog.

10% Credible Interval

Mean
90% Credible Interval
CV (%)
? Because

QT c0
%
41.8
98.6
403.1
166.6

%
27.7
41.7
74.8
153.2

BSV
A
%
30.7
48.6
96.0
229.1

%
26.7
46.4
91.2
177.1

Slope?
 ms 
nM

2.0105
3.8105
7.2105
164.9

IOV
QT c0
[ms]
10.71
13.45
17.50
43.0

ISV

[ms]
9.26
9.42
9.59
1.4

of Slope BSV 10% credible interval is negative, Slope BSV has not been represented via coefficient of variation
since such measure is not definable for negative quantities.

Table 3.8: Mean, 80% credible intervals and precision (expressed as percentage coefficient
of variation) of the random effects of the PKPD model of equation (2.12) for dog data.
Unlike standard ways of estimation, Basyesian estimation handles every enitity as a
distribution, hence posteriors and credible intervals may be obtained also for parameters

3.2. PHARMACODYNAMIC ANALYSIS

39

conveying variability in the model, i.e. (Between Subject Variability), 2 (Inter Occasion
Variability) and 2 (Intra Subject Variability). Table 3.8 contains MMSE estimator and
80% credible interval of parameters that quantify each source of variability. It may be
noticed that CV values in BSV are all above 100%, reflecting an inaccurate estimate on
how much the parameters vary amongst individuals. Such weakness in detecting reliable
measures of the BSV is due to the very low number of dogs (4) enrolled in the PD study.

3.2.3

Clinical Data

Since there is no difference in the PKPD model between species, the same kind of dissection
performed for preclinical data in section 3.2.2 will be presented. Starting from the estimated
posterior distributions, Figure 3.9 shows profiles of A and are too wavy to conclude
Markov chains for those parameters have fully convegred (Figure 3.9c and 3.9d), reinforcing
the hypothesis made in section 3.2.1 that the number of QT measurements collected per
day is too little to be able to extrapolate information on the cyrcadian rhythm.

10% Credible Interval

Mean
90% Credible Interval
CV (%)

QT c0
[ms]
389.7
392.9
396.1
0.6

unitless
0.201
0.239
0.275
12.2

A
[ms]
2.77
3.44
4.09
23.0

[hr]
5.97
7.33
8.59
27.1

Slope
ms
[ nM
]
1.6106
4.2106
7.1106
52.9

Table 3.9: Mean, 80% credible intervals and precision (expressed as percentage coefficient
of variation) of the population parameters of the PKPD model of equation (2.12) for clinical
data.

10% Credible Interval

Mean
90% Credible Interval
CV (%)

QT c0

30.1
44.2
81.5
47.7

27.9
38.9
58.1
29.4

BSV
A
%
29.0
44.9
69.2
35.0

Slope

31.9
63.2
104.4
45.6

61.8
113.2
292.0
37.8

IOV
ISV
QT c0

[ms]
15.2 12.66
17.5 13.27
20.3 13.94
23.2
3.8

Table 3.10: Mean, 80% credible intervals and precision (expressed as percentage coefficient
of variation) of the random effects of the PKPD model of equation (2.12) for clinical data.
With regards to the MMSE estimates of the models parameters, it has to be mentioned
that results from human data lead to the same conclusions in terms of drug effect. Indeed,

40

CHAPTER 3. RESULTS

15

10
0.10

pdf()

pdf(QTc0)

0.15

0.05

0.00

0
385

390

395

400

405

0.15

QTc0[ms]

0.20

(a)

0.25

0.30

0.35

(b)

0.8
0.4

0.3

pdf()

pdf(A)

0.6

0.4

0.2

0.2

0.1

0.0

0.0
2

10

12

[hr]

A [ms]

(c)

(d)

pdf(Slope)

150000

100000

50000

0
0.000005

0.000005

0.000015

Slope [ms/Nm]

(e)

Figure 3.9: Posterior distributions of the PKPD model parameters in human data: QT c0
(a), (b), A (c), (d) and Slope (e).

3.2. PHARMACODYNAMIC ANALYSIS

41

as reported in Table 3.9, the Slope estimate is again (approximately) zero. Further, Table
3.10 confirms that with a reasonable number of individuals (45) estimates of BSV are
sufficiently accurate (CVs are less than 50%). Nevertheless, because of the MMSE estimate
of Slope is close to zero, the estimated BSV of this parameter is quite high; however, since
it can be deduced the influence of drug effect in observed data is absent, there is no interest
in the knolwedge of how Slope parameter varies among individuals.

Chapter 4

Conclusions
Concerning the assessment of risky QTc increases, according to regulatory guidelines [4] a
threshold of 10 msec was used to explore the probability of QTc interval prolongation at
the relevant therapeutic range. In particular, Table 4.1 enables to infer GSK945237 does
not prolong QTc interval in either humans or dogs since the concentration needed to have
a 50% probability of a QTc prolongation grater than 10 msec is approximately hundred
times bigger than the predicted Cmax .

DOG
HUMAN
? concentration

?
P C50
[nM]
3500000
4200000

Cmax
[nM]
62936.18
13096.86

associated with a P(QTc increase 10 msec) = 50%

Table 4.1: P C50 estimates and predicted Cmax in both preclinical and clinical data.
These results differed from typically reported results in telemetered dogs, which are
often based on non-parametric methods and statistical summaries of the data; yet, unlike traditional, data-driven experiments whose outcomes are often vague and ambigous,
a model based approach provides a robust method to properly evaluate cardiovascular
safety studies, allowing to advance clear and define conclusions. Moreover, the flexibility
of the proposed approach, characterised by the usage of a Bayesian framework, eases and
optimizes the interpretation of the results, as proven by the opportunity of obtaining the
probability of a QTc prolongation greater than 10 msec (Figure 4.1). In fact, the assessment of a probability measure rather than a yes or no answer to the question: is the
QTc prolongation provoked by the drug under investigation matter of concern for further
development of the molecule? gives considerable advantages in a decision analysis context.

43

44

CHAPTER 4. CONCLUSIONS

P(QTc prolongation >= 10ms)

0.8

0.6

0.4

0.2

0.0
0

5000000

10000000

Concentration [nM]

Figure 4.1: Curves showing the relationship between concentration and the probability of
a QTc increase grater than 10 msec in dogs (gray line) and humans (black line).
Although no QTc-prolonging effects were observed, results illustrate the value of a quantitative approach to characterise drug effects in early drug development. Furthermore, the
analysis reconfirms that accurate interpretation of pre-clinical findings requires suitable
pharmacokinetic sampling and some understanding of expected therapeutic exposure.
In the context of translational pharmacology, the proposed model showed it is able to
predict both human and dog data (see Figure 4.2) making its use suitable for interspecies
scaling of QT interval prolongation. For this purpose, performing the presented analysis for
an adeqaute number of compounds known to have QT prolonging effects may give essential
information on the relationship of such drug effect between humans and dogs. In other
words, estimation of Slope parameter in several drugs and in both species would allow to
find the existing correlation of that parameter between human and dogs, bridging the gap
between the two species and supporting decision making analysis at the very beginning of
the development of a new compund.

45

Individual Predicted QT [ms]

500

400

300

200

200

300

400

500

Observed QT [ms]

Figure 4.2: Individual goodness of fit plot of the PKPD model for both preclinical (gray
dots) and clinical (black dots) data; the solid line represents the unity line. The model is
capable to predict both dog and human QT interval.

Bibliography
[1] J. Milnes, O. Crociani, A. Arcangeli, J. Hancox, and H. Witchel, Blockade of HERG
Potassium Currents by Fluvoxamine: Incomplete Attenuation by S6 Mutations at
F656 or Y652, British Journal of Pharmacology, vol. 139, pp. 887898, 2003.
[2] J. Valentin, Reducing QT Liability and Proarrhythmic Risk in Drug Discovery and
Development, British Journal of Pharmacology, vol. 159, pp. 511, 2010.
[3] R. Fenichel, M. Malik, C. Antzelevitch, M. Sanguinetti, D. Roden, S. Priori, J. Ruskin,
R. Lipicky, and L. Cantilena, Drug-induced Torsades de Pointes and Implications for
Drug Development, Journal of Cardiovascular Elettrophysiology, vol. 15, pp. 475495,
2004.
[4] Guidance for Industry: ICH E14 Clinical Evaluation of QT/QTc Interval
Prolongation and Proarrythmic Potential for Non-Antiarrythmic Drugs, 2005.
Department of Health and Human Services, Food and Drug Administration.
http://www.fda.gov/downloads/RegulatoryInformation/Guidances/ucm129357.pdf.
[5] A. Chain, K. Krudys, M. Danhof, and O. Della Pasqua, Assessing the Probability of Drug-Induced QTc-Interval Prolongation During Clinical Drug Development,
Clinical Pharmacology & Therapeutics, 2011.
[6] D. Lunn, A. Thomas, N. Best, and D. Spiegelhalter, WinBUGS a Bayesian modelling framework: concepts, structure, and extensibility, Cambridge, UK: MRC Biostatistic Unit, vol. 10, pp. 325337, 2000.
[7] S. Beal, L. Sheiner, A. Boeckmann, and R. Bauer, NONMEM Users Guides (19892009). Icon Development Solutions, Ellicott City, MD, USA, 2009.
[8] J. Gabrielsson and D. Weiner, PK/PD Data Analysis: Concepts and Applications.
Stockholm, Sweden: Swedish Pharmaceutical Press, 2007.
[9] H. Bazett, An analysis of the time relations of electrocardiograms, Heart, vol. 7,
pp. 353370, 1920.
47

48

BIBLIOGRAPHY

[10] L. Fridericia, Die systolendauer im elektrokardiogramm bei normalen Menchen und

bei Herzkranken, Acta Med Scand, vol. 53, pp. 469486, 1920.
[11] W. Nelson, Y. Tong, J. Lee, and F. Halberg, Methods for cosinorrhythmometry,
Chronobiologia, vol. 6, pp. 305323, 1979.
[12] Y. Touitou and E. Haus, Biologic Rhythms in Clinical and Laboratory Medicine. Berlin,
Germany: Springer Verlag.
[13] K. Murphy, Conjugate Bayesian Analysis of the Gaussian Distribution,
2007.
The Univeristy of Bristish Columbia. http://www.cs.ubc.ca/ murphyk/Papers/bayesGauss.pdf.
[14] Jaws: repeated measures analysis of variance. WinBUGS Help Example on the Use
of Wishart Distribution.
[15] M. O. Karlsson and N. Holford, A tutorial on visual predictive check, 2008. Seventeenth PAGE meeting, Abstract 1434. www.page-meeting.org/?abstract=1434.
[16] J. Geweke, Evaluating the accuracy of sampling-based approaches to calculating
posterior moments, Bayesian Statistics, vol. 4, pp. 169194, 1992.
[17] A. Gelman and D. Rubin, Inference from iterative simulation using multiple sequences, Statistical Science, vol. 7, pp. 457511, 1992.
[18] A. Raftery and S. Lewis, How many iterations in the Gibbs sampler?, Bayesian
Statistics, vol. 4, pp. 763774, 1992.
[19] P. Heidelberger and P. Welch, Simulation run length control in the presence of an
initial transient, Operations Research, vol. 31, pp. 11091144, 1992.
[20] I. Ntzoufras, Bayesian Modeling Using WinBUGS. Hoboken, NJ, USA: John Wiley
& Sons, Inc., 2008.